XFT Data Volume Studies + Possible Solutions

1. XFTData Volume Studies +Possible Solutions Alison Lister Robin Erbacher, Rob Forrest, Andrew Ivanov, Aron Soha Alison Lister, UC Davis

2. Large Data Volume • Current limitations of combined system • Ran fine for whole store with initial Luminosity 190e30 • Data corruption when initial Luminosity 240e30 • Both XFT alone and L2CAL alone are known to work fine up to 240e30 • Hypothesis • Data gets corrupted due to high load on PCI bus and/or buffer sizes of input FIFO in FILAR channel • Solutions to reduce load on PC • Split to more XFT inputs (6 XFT configuration) • Use more than 1 PC for inputs • Reduce total size of XFT data (firmware changes) Alison Lister, UC Davis

3. Mean of the data size should not be a problem Tails are likely to be the issue Higher luminosity -> More frequently several high occupancy events in a row Including all highest luminosity runs to date from dstream data ~100’000 events in total Mean progresses ~linearly Tails also seem to progress linearly Total XFT Size Lumi ~350e30: ~800 words Alison Lister, UC Davis

4. 6 XFT Implementation • Splitting the XFT data into 6 inputs to the PC will reduce data size per FILAR input • Easier to balance the loads between PCI buses • Total data size to PC remains the same • Hardware and firmware validated and working fine • 6th Pulsar board validated since last Thursday • FER code changes to enable/disable Pulsar input channels from database • set via CardEditor • All PC code changes implemented and tested in beam and between stores Alison Lister, UC Davis

5. 6 XFT Tests • If still use 1 PC, the old inputs (not SVT) need to be merged • Merged code written and tested • Tested in beam with 6 XFT and merged old inputs up to Lumi ~200e30 • Tested at EOS with 6 XFT, 4 L2CAL and merged old inputs • Still need to investigate additional latency for merging old inputs • Preliminary studies show that tails in latency are significantly increased • Tails up to ~20us when using 6XFT + merged old input configuration Alison Lister, UC Davis

6. 6 XFT Tests • If use 2 PCs • Many alternatives presented in a previous talk • 2nd PC as “fast abort” for XFT • 2nd PC doing all L2CAL related algorithms and transferring output of clustering to decision PC • … • Would remove the problem not just push it back • Control nodes (designed for ATLAS) are designed for multi-PC processing farms • Can’t investigate further until have more PCs • Especially if moving to driving with 1 of the new PCs Alison Lister, UC Davis

7. 2 PC “Proof of Principle” • Copied code to 2nd new PC • Enabled only 1 channel • Copied input to Solar output • Worked first try! • All commands (e.g. HRR) passed on to both PCs in control node • Data always in sync • No problems seen in 1/2h running at EOS Configuration using 2 PCs is in principle possible • Use 2nd PC as a “pass through” for XFT1 • Added 2nd PC to PULSAR_TED control node L2TS Decision PC: Old algo. + Clustering algo. L2CAL Old XFT2+3 XFT 1 Simple copy of input to output XFT 1 Alison Lister, UC Davis

8. Studies Related to Reducing Total XFT Data Size Alison Lister, UC Davis

9. Reducing Data Size: Removing SL3 Cot(theta) Resolution It has been suggested that removing SL3 might help with the data volume • SL3 highest occupancy • High lumi might be all firing anyway • Investigated the degradation to the resolution at low lumi • Only have 2 points not 3 so always a straight line • Removing 1 powerful fake rejection at lower luminosities Using 3 SL 0.12 Lumi 80-90e30. High quality muons • Cleaner than data ever could be! • Cot(theta) resolution 3 times worse • Z0 resolution ~4 times worse • # hit combinations per axial track increases dramatically (see next slide) Z0 Resolution Using 3 SL 10cm Alison Lister, UC Davis

10. Reducing Data Size: -SL3 • Degeneracy in possible stereo tracks when including SL3 or not • 0 degeneracy = 1 possibility • > 0 degeneracy: have to randomly pick 1 combination • Luminosity dependence being investigated…. • Note : Clean environment! Real data will look much worse Alison Lister, UC Davis

11. Regions of Interest • Concept is the following • Only the information related to regions containing tracks that L2XFT algorithms could be interested in are sent to the PC • These regions are to be defined by L1 bits for example • “Uninteresting” regions send empty packets • Requires firmware changes in Pulsar board • Need L1 bits as inputs • Need fast abort infrastructure when no regions of interest present • Stereo reconstruction algorithm written such that all cells within 3 cells of the expected hit are considered • Consider all finders that need to be read out in order to get the whole region around the track to be confirmed Alison Lister, UC Davis

12. Regions of Interest • Would need to implement a fast-abort based on L1 trigger bits • Only send to PC data from finders located within 3 cells of axial track • Non-linear increase in mean size after sparsification • Prereq: L1 CMX trigger, track pT > 4 GeV • Tails just as large as with no ROI Tails remain Alison Lister, UC Davis

13. Conclusions • XFT with 3 inputs well well for current luminosities without integrated system • Many possible solutions to the combined XFT + L2CAL operational problems at high luminosity • Would like a quick temporary solution that has a chance of working at high lumi • e.g the 6 XFT + L2CAL + Merged old if additional latency not an issue • Available for Trigger Fest before Shutdown so as not to slow down algorithm development • For the future we also need a proper solution that is guarantied to work at high lumi • e.g. develop in parallel the 2 PC solution • Available as soon as possible with minimal extra time-investment • These temporary and long-term solutions could be one and the same or slight variations of each other Alison Lister, UC Davis

14. Backup Alison Lister, UC Davis

15. SL3 SL5 SL7 sizes Alison Lister, UC Davis